360 likes | 449 Views
Differential (De)Serialization for Optimized SOAP Performance. Michael J. Lewis Grid Computing Research Laboratory Department of Computer Science Binghamton University State University of New York (with Nayef Abu-Ghazaleh, Madhu Govindaraju ). Motivation.
E N D
Differential (De)Serialization for Optimized SOAP Performance Michael J. Lewis Grid Computing Research Laboratory Department of Computer Science Binghamton University State University of New York (with Nayef Abu-Ghazaleh, Madhu Govindaraju)
Motivation • SOAP is an XML-based protocol for Web Services that (usually) runs over HTTP • Advantages • extensible, language and platform independent, simple, robust, expressive, and interoperable • The adoption of Web Services standards for Grid computing requires high performance
The SOAP Bottleneck Serialization and deserialization • The in memory representation for data must be converted to ASCII and embedded within XML • Serialization and deserialization conversion routines can account for 90% of end-to-end time for a SOAP RPC call [HPDC 2002, Chiu et. al.] • Our approach • Avoid serialization and deserialization altogether, whenever possible • bSOAP: Binghamton’s SOAP implementation
Overview of the Optimizations • Differential Serialization (DS) (sender side) • Save a copy of the last outgoing message • If the next call’s message would be similar, then • use the previous message as a template • only serialize the differences from the last message • Differential Deserialization (DDS) (receiver side) • Checkpoint incoming message portions • If the next incoming message is similar, then • use the deserialized values from the last message • only deserialize the differences from the last message
DS and DDS • DS and DDS are separate, different, disjoint optimization techniques • sender side (DS) vs. receiver side (DDS) • data update tracking (DS) vs. parser checkpointing (DDS) • neither depends on the other • each takes advantage of sequences of similar messages to avoid expensive SOAP message processing • neither changes the protocol, what goes in the SOAP message, or on the wire • each remains interoperable with other SOAP implementations
DS: Update Tracking • How do we know if the data in the next message will be the same as in the previous one? • If it is different, how do we know which parts must be reserialized? • How can we ensure that reserialization of message parts does not corrupt other portions of the message?
Data Update Tracking (DUT) Table struct MIO { int a; int b; double val;}; int mioArray(MIO[] mios) Field TPointer SLength FWidth Dirty? X 5 5 YES Y3 7 YES Z 5 10 NO POST /mioExample HTTP/1.1 . <?xml version='1.0'?><SOAP-ENV:Envelope ...> . <x xsi:type='xsd:int'>12345</x> <y xsi:type='xsd:int'>678</y> <z xsi:type='xsd:double'>1.166</val> . </SOAP-ENV:Envelope>
Problems and Approaches • Problems • Some fields require reserialization • The current field width may be too small for the next value • The current message (or chunk) size may be too small Solving these problems enables DS, but incurs overhead • Approaches • shifting • chunking • stuffing • stealing • chunk overlaying
Shifting • Shifting: Expand the message on-the-fly when the serialized form of a new value exceeds its field width • Shift the bytes of the template message to make room • Update DUT table entries for all shifted data …</w><x xsi:type='xsd:int'>1.2</x><y xsi:type=…. becomes …</w><x xsi:type='xsd:int'>1.23456</x><y xsi:type=…. • Performance penalty • DUT table updating, memory moves, possible memory reallocation
Stuffing • Stuffing: Allocate more space than necessary for a data element • explicitly when the template is first created, or after serializing a value that requires less space • Helps avoid shifting altogether • Doesn’t work for strings, base64 encoding …<y xsi:type='xsd:int'>678</y><z xsi:type=… can be represented as …<y xsi:type='xsd:int'>678</y><z xsi:type=…
Stealing • Stealing: Take space from nearby stuffed fields • Can be less costly than shifting [ISWS ‘04] …'>678</y><z xsi:type='xsd:double'>1.166</val> y can steal from z to yield… …'>677.345</y><z xsi:type='xsd:double'>1.166</val> • Performance depends on several factors • Halting Criteria: When to stop stealing? • Direction: Left, right, or back-and-forth?
Performance • Performance depends on • which techniques are invoked • “how different” the next message is • Message Content Matches • identical messages, no dirty bits • Perfect Structural Matches • data elements and their sizes persist • Partial Structural Matches • some data elements change size • requires shifting, stealing, stuffing, etc. • We study the performance of all our techniques on synthetic workloads of scientific data • summary: 17% 10X improvement
Perfect Structural Matches • Perfect Structural Matches: • Some data items must be overwritten (DUT table dirty bits) • No shifting required • Performance study: • vary the message size • vary the reserialization percentage • vary the data type • Doubles and • Message Interface Objects (MIO’s, <int, int, double>) (not shown)
Send Time depends directly on % serialized • Important to avoid reserializing
DDS: The Approach • As an incoming message is being processed • store parser states periodically • compute corresponding message portion checksums • For subsequent (hopefully similar) messages • compare incoming message checksums with stored checksums (fast mode parsing) • if checksums match • the parser can skip to the next parser state without actually generating it from the incoming message • on checksum mismatch • revert to regular mode parsing
Effectiveness • Depends on… • Similarity in consecutive messages • determines how often in fast mode • How much faster fast mode is • deserialization vs. checksum calc and compare • Efficiency in identifying mode switches • Checkpoint and checksum overhead
Creating Checkpoints • First one right after start tag that contains the name of the back end element • Thereafter, checkpoints are created periodically • based on number of bytes processed • configurable parameter of bSOAP • Tradeoff: overhead vs. fast mode processing time • standard implementation: full parser checkpoints • optimization [Grid 2005]: differential checkpoints
Fast mode parsing • Parser reads messages and computes checksums on message portion boundaries • a match allows the parser to “skip” to the next saved state • Switching back to fast mode • must compare current and stored parser states • matching stack contains necessary structural info • namespace aliases must also be the same • stored and checked separately
Performance Summary • Without DDS: comparable to gSOAP • DDS Overhead • message portions 256 4K; overhead < 10% • message portion size 32 bytes too small • DDS improvement • large “hard to deserialize” messages, very similar 25X speedup (upper bound) • Dual mode performance • depends on message portion size, where and how often mode switches take place • can reduce by a factor of 3, or be slightly slower
Benchmark Suite for SOAP-Based Grid Web Services • Motivation • Web services based applications have diverse requirements • SOAP and XML present design and implementation challenges • Several novel efforts exist to address key bottlenecks • examples can be found in gSOAP, bSOAP, .NET • A benchmark suite • can help determine the best available toolkit • based on communication patterns and data structures in use • Benchmarks and performance evaluation framework • Drivers, WSDL files and Java code • helps provide insights on opportunities for optimization • Madhu Govindaraju: mgovinda at cs.binghamton.edu • http:// grid.cs.binghamton.edu / projects / soap_bench
Thank You • For More information • Grid Computing Research Laboratory • SUNY – Binghamton: Computer Science Department • http:// grid.cs.binghamton.edu • mlewis at binghamton.edu • DS: [HPDC ’04], [IC ‘04] • DDS: [SC ’05], [Grid2005]
Experimental Setup • Machines • Dual Pentium 4 Xeon 2.0 GHz, 1 GB DDR RAM, 15K RPM 18 GB Ultra-160 SCSI drive. • Network • Gigabit Ethernet. • OS • Debian Linux. Kernel version 2.4.24. • SOAP implementations • bSOAP and gSOAP v2.4 compiled with gcc version 2.95.4, flags: -O2 • XSOAP 1.2.28-RC1 compiled with JDK 1.4.2 • bSOAP/gSOAP socket options: SO_KEEPALIVE, TCP_NODELAY,SO_SNDBUF = SO_RCVBUF = 32768 • Dummy SOAP Server (no deserialization).
Message Content Matches • Message Content Match: • The entire stored message template can be reused without change • No dirty bits in the DUT table • Best case performance improvement • Performance Study • compare gSOAP, XSOAP, and bSOAP, with differential serialization on and off • vary the message size • vary the data type: doubles and MIO’s (not shown)
bSOAP ~= gSOAP • 10X imprvmt in DS • (expected result) • Upper bound
Shifting • Partial Structural Match: • Not all of array elements are reserialized • Performance Study • Intermediate size values to maximum size values. • Array of doubles (18 24) • Array of MIO’s (36 46)(not shown)
100% 75%: Imprvt 23% 75% 50%: Imprvt 31% 50% 25%: Imprvt 46%
Stuffing • Closing Tag Shift: • Stuffed whitespace comes after the closing tag • Must move the tag to accommodate smaller values • Performance Study • send smallest values (1 char) • vary field size: smallest, intermediate, maximum • Array of doubles (max = 24, intermediate = 18, min = 1) • Array of MIOs • (max = 46, intermediate = 38, min = 3) (not shown)
Closing tag shift, not increased message size, effects stuffing performance
Summary • SOAP performance is poor, due to serialization and deserialization • Differential serialization • Save a copy of outgoing messages, and serialize changes only, to avoid the observed SOAP bottleneck • Techniques: • Shifting, chunking, chunk padding, stuffing, stealing, chunk overlaying • Performance is promising (17% to 10X improvement), depends on similarity of messages
Other Approaches • SOAP performance improvements • Compression • Base-64 encoding • External encoding: Attachments (SwA), DIME • These approaches may be necessary and can be effective. However • they undermine SOAP’s beneficial characteristics • interoperability suffers • The goal • improve performance, retain SOAP’s benefits
Applications that can Benefit • Differential Serialization is only beneficial for applications that repeatedly resend similar messages • Such applications do exist: • Linear system analyzers • Resource information dissemination systems • Google & Amazon query responses • etc.
Data Update Tracking (DUT) Table • Each saved message has its own DUT table • Each data element in the message has its own DUT table entry, which contains: • Location: A pointer to the data item’s current location in the template message • Type: A pointer to a data structure that contains information about the data item's type. • Serialized Length: The number of characters needed to store the last written value • Field Width: The number of allocated characters in the template • A Dirty Bit indicates whether the data item has been changed since the template value was written
Updating the DUT Table • DUT table dirty bits must be updated whenever in-memory data changes • Current implementation • explicit programmer calls whenever data changes • Eventual intended implementation • more automatic • variables are registered with our bSOAP library • data will have accessor functions through which changes must be made • when data is written, the DUT table dirty bits can be updated accordingly • disallows “back door” pointer-based updates • requires calling the client stub with the same input param variables
Worst Case Shifting • “Worst case shifting”: • All values are reserialized from smallest size values to largest size values. • Performance Study • vary the chunk size (8K and 32K) • Array of doubles (1 24). • Array of MIOs (3 46) (not shown)
Worst case shifting is 4X slower Reducing chunk size doesn’t help